Large Language Models Acting as Psychotherapists: A Scoping Review

As LLMs are introduced as "therapists," HCI research has begun to reconcile the tension between controllability (manualized, safer flows) and complexity (open‑ended, higher‑risk dialogue). We review 76 studies of LLMs‑as‑therapist using the NIH Stage Model, synthesizing psychotherapeutic characteristics, adaptation methods and datasets, and evaluation practices. Most systems are CBT‑centric or unspecified; interventions are single‑session with general well‑being targets. Implementations rely on fine‑tuning/prompting with retrieval and sensing; among 23 datasets, synthetic corpora are larger yet far shorter than real conversations. Evaluations are heterogeneous, frequently using bespoke tools or LLM‑as‑judge; safety, privacy, and diversity are rarely measured. We contribute a human‑centered map of this space, a taxonomy of evaluation constructs, and an HCI‑oriented agenda: scaffold conversations to reflect clinical processes, implement explicit crisis pathways and guardrails, use validated and user‑centered measures, and adopt equity‑aware data and reporting practices. We position LLMs‑as‑therapist as a core interactive systems challenge for HCI.